Type 1 diabetes is a chronic disease, which presently cannot be prevented or cured. It is treated with insulin therapy and actively managed through blood glucose (BG) control. To avoid serious diabetic complications, patients must monitor their BG levels throughout the day, striving to avoid both hyperglycemia (high BG levels) and hypoglycemia (low BG levels). While continuous glucose monitoring (CGM) sensors and insulin pumps with ?exible dosing may aid in achieving good BG control, management of diabetes is still dif?cult and laborious for patients and physicians. It is complicated by a wide variability among individual patients in terms of physiological responses to treatment as well as to life events such as stress, exercise, or changes in schedule and sleep. New portable sensing technologies have been recently developed for providing almost continuous measure- ments of an array of physiological parameters that include heart rate, skin conductance, skin temperature, and properties of body movements such as acceleration. The main research objective of this project is to leverage data acquired from wearable physiological sensors to build accurate, personalized blood glucose level prediction models for diabetes management. Predicting BG control problems before they occur would give patients time to intervene and prevent these problems. This would enhance patient safety and contribute to improved overall control, with its concomitant reduction in costly complications. Blood glucose level prediction is very complex problem. Recent advances in unsupervised feature learning and deep learning have made it possible to learn complex models from data using simple algorithms. Inspired by these signi?cant developments in Arti?cial Intelligence (AI), we propose to employ unsupervised feature learn- ing and deep learning techniques in order to build an architecture for modeling blood glucose behavior that can seamlessly incorporate data coming from any number of physiological sensors. A recurrent neu- ral network (RNN) will be trained to capture dependencies among the input physiological parameters that are relevant to BG prediction. To account for individual patient differences, a predictive model will be developed for each patient by training on the features discovered by the RNN. The primary impact of this work would be to improve the overall health and quality of life for the 1.25 million Americans with type 1 diabetes. Accurate prediction models would enable practical applications ranging from alerts of impending problems to decision support tools for evaluating the effects of different food or lifestyle choice. Additionally, the wealth of patient and sensor data collected for this work will lead to the creatin of de-identi?ed datasets to be used by the research community in evaluating approaches to blood glucose prediction. The new approaches developed for this complex domain may aid in the development of time series forecasting models for a broad array of sensor-enabled applications in other health and wellness domains.
To aid in diabetes management, machine learning models will be built to predict future blood glucose levels based on wearable sensor data from commercially available fitness bands in addition to blood glucose, insulin and meal data. These models could help the over one million Americans with type 1 diabetes to anticipate and prevent blood glucose control problems before they occur. This would enhance patient safety and contribute to improved overall blood glucose control, with its associated reduction in costly complications.
| Schwartz, Frank L; Marling, Cynthia R; Bunescu, Razvan C (2018) The Promise and Perils of Wearable Physiological Sensors for Diabetes Management. J Diabetes Sci Technol 12:587-591 |
